BIDU (Trans): Cloud margins still have room to climb, yet we hold firm in betting on AI.

Below is Dolphin Research's Trans of $ Baidu.US FY26Q2 earnings call.

I. Key Takeaways

1. 2H Guidance

a. AI cloud infra: based on current demand, the customer pipeline, and our full‑stack differentiation, revenue is set to remain strong in 2H and could even accelerate.b. Ads: the AI search transition, along with our deliberate restraint on monetization and pricing, has been a drag and is likely to persist, keeping ads under pressure in 2H.c. Apollo Go: priorities are safety and ops excellence, global expansion, scaling fleet and orders, and pushing more cities toward unit‑economics breakeven.

2. Capital Markets

a. Dual primary listing in HK: the BOD approved the voluntary conversion in Jul, application was submitted and accepted by HKEX; an EGM on Aug 26 will seek shareholder approvals, and effectiveness is expected within the year (subject to shareholder and HKEX approvals). Post‑conversion we expect a broader investor base, better liquidity, and enhanced financing flexibility across markets.b. Southbound Stock Connect: after conversion, we are preparing for potential inclusion as early as possible, subject to eligibility and exchange reviews.c. Kunlun Chip spin‑off listing is progressing; management will update the market when there is material progress.

3. Key Financials This Quarter

a. Topline: revenue RMB 31.3bn (-4% YoY, -2% QoQ); Baidu Core RMB 25.2bn (-4% YoY, -3% QoQ); iQIYI RMB 6.3bn (-5% YoY, +1% QoQ).b. Segment: AI‑related revenue in Baidu Core RMB 12.5bn, still about half of Core; AI cloud infra +~50% YoY, with GPU cloud +283% YoY.c. Costs & opex: COGS RMB 19.1bn (-3% QoQ on lower AI cloud‑related costs, partly offset by higher TAC; +4% YoY on higher AI cloud costs); opex RMB 9.2bn (-1% QoQ, -17% YoY), with the YoY decline mainly from lower ECL, channel spend, and R&D personnel‑related costs.

d. Profit: OP RMB 3.0bn, OPM 10%; non‑GAAP OP RMB 3.8bn, OPM 12%; attributable net income RMB 2.3bn, NPM 7%, diluted EPS per ADS RMB 5.74; non‑GAAP attributable net income RMB 2.6bn, NPM 8%, non‑GAAP diluted EPS per ADS RMB 7.22.e. Non‑op & one‑offs: other income net RMB 184mn vs. RMB 626mn in Q1 and RMB 4.9bn YoY, with the YoY drop mainly from lower FV gains on LT investments and higher net FX losses from RMB/USD moves; income tax expense RMB 1.0bn vs. RMB 528mn in Q1 and RMB 881mn YoY.

f. Cash & headcount: as of Jun 30, 2026, cash and investments totaled RMB 283.1bn (cash & equivalents, restricted cash, ST investments, net LT time deposits & HTM, and adj. LT investments). Operating cash flow RMB 3.4bn; Baidu Core headcount ~27k.

4. Margin Trend & Investment Discipline

a. AI cloud infra delivered rapid profit growth with YoY margin expansion alongside high revenue growth, with mix shifting toward higher‑quality GPU cloud.b. We remain in the AI investment cycle with unchanged commitment, but stress demand‑driven investment by customers and internal biz, with most spend put to work quickly and contributing revenue earlier. We continue to strengthen supply‑chain management to lift capital efficiency, keeping discipline on ROIC, operating efficiency, and cash flow.

II. Details from the Call

2.1 Management Remarks

1. In‑house AI chips (Kunlun Chip)

a. Demand stayed strong and broadened across industries, with more customers adopting our chips and a wider set of AI workloads supported.b. The software stack kept improving, expanding compatibility with mainstream models/frameworks and easing enterprise deployments; beyond ERNIE and other leading domestic FMs, we added coverage for Kimi, GLM 5.2, and MiniMax M3 this quarter.c. Inference throughput and overall compute efficiency improved, enabling more diverse and demanding AI workloads.d. Over the past decade we have developed and commercialized three chip generations; the roadmap is clear: the latest M100 optimized for large‑scale inference has launched, and the next‑gen M300 is imminent.

2. AI Cloud Infra & Qianfan (MaaS)

a. AI cloud infra revenue rose ~50% YoY, outgrowing the industry; training and inference demand remained robust while market compute supply stayed tight.b. Usage and spend from existing key customers (incl. top players in online gaming, e‑commerce, and lifestyle content) kept rising, while total customer count grew rapidly across company sizes.c. Demand spans Internet, embodied AI, AV, smartphones, and financial services; Internet and AV remained strong, and embodied AI revenue grew ~6x YoY.d. GPU cloud revenue +283% YoY, the fourth straight triple‑digit quarter and an acceleration from Q1's 184%, with a rising mix within AI cloud infra; given its superior margin structure, this mix shift lifts overall revenue quality and long‑term profitability.

e. Qianfan's model hub covers ERNIE and nearly all major domestic models; inference optimization improved throughput and stability while lowering latency and cost, driving >9x YoY revenue growth from external customer token usage this quarter, led by rapid increases in daily token consumption.

3. ERNIE Foundation Model

a. We reorganized the model team into foundation model and application groups with clearer mandates, and recently onboarded top‑tier AI talent into foundation research.b. Investment in foundation innovation remains intact, supporting continuous ERNIE upgrades and underpinning innovation at both the model and application layers.

4. AI Apps (Digital Human, Miaoda, Pharmo, General‑purpose Agents)

a. Digital human performance improved and costs fell, expanding use cases from e‑commerce livestreaming and avatar videos to real‑time interactive avatars and newly launched video podcasts.b. We signed multiple leading clients and saw clear scale‑up among existing ones — a well‑known China Internet company expanded digital‑human livestreaming to ~2.5x after one‑quarter pilot; the overseas platform launched last quarter is progressing well.c. Miaoda 3.0 lets users build standalone Android/iOS apps via natural language, even on phones; users shifted from one‑off trials to repeat iterations, lifting stickiness. Jun MAU rose 67% vs. Mar, with coverage extending from tech/education to healthcare, manufacturing, finance, and logistics.

d. Pharmo Agent 2.0 kept improving usability and expanding scenarios, attracted leading enterprises, and achieved early commercialization this quarter, with value extending from efficiency to ops optimization.e. Doomate, our general‑purpose agent (launched earlier this year across PC and mobile), rolled out an enterprise edition and kept expanding proprietary skills and toolsets.f. Baidu Wenku and Baidu Netdisk are fully AI‑enabled, and GenFlow was upgraded this quarter; the AI feature penetration across the two rose 27.4% YoY in Jun.

5. AI Search & ERNIE Assistant

a. AI‑generated answers kept improving in reliability, structure, and presentation; hallucination stayed low, and model judgment of content quality strengthened, reducing low‑quality answers and enhancing UX and satisfaction.b. AI Search further integrated with ERNIE Assistant, turning one‑off answers into interactive, multi‑turn conversations.c. In Jun, ERNIE Assistant DAU rose 83% YoY, with daily conversation rounds up more than 3x YoY.

6. Apollo Go (Robotaxi)

a. Hong Kong: obtained the first fully driverless test permit in Jun and began tests on the airport island in Jul, making us the first to run fully driverless tests in a right‑hand‑drive, left‑lane market. These learnings accelerated progress in London, where we started open‑road tests in Jul with Uber and Lyft.b. Other overseas: launched fully driverless commercial ops in Dubai in Jul, now the largest among local driverless services, available via Apollo Go and Uber apps; started open‑road tests in Switzerland with a local partner (transcript 'Host', likely PostAuto); signed an MoU with Kazakhstan's Private Holding Limited to explore local Robotaxi services.c. Orders: ~1.0mn fully driverless orders fulfilled in Q2; as of Jun 2026, cumulative rides to the public exceeded 23.0mn. Q2 orders were temporarily affected by operational adjustments in some China cities due to regulatory considerations; we conducted a systematic review to enhance system robustness and ops rigor, and affected cities began resuming operations from Aug.

d. Safety & UX: as of end‑Jun, on average a fully driverless vehicle had one airbag deployment per ~14.4mn km; we improved pick‑up/drop‑off recommendations to shorten walking distance and avoid unsuitable stops, and upgraded perception and motion planning for a smoother ride.

2.2 Q&A

Q: With trillion‑parameter models rapidly emerging and resetting benchmarks, how do you view ERNIE's competitive position? After bringing in top foundation‑model talent, what are the tech/product priorities, and what should investors expect next?

A: Foundation models are evolving fast, with different leaders on specific capabilities every few months, underscoring a highly dynamic market and an unsettled competitive landscape. In such markets, sustained tech investment, application‑driven methodology, and patience define long‑term competitiveness — BIDU has long believed in tech and invested for the long haul, and meaningful innovation repeatedly proves to require persistence.Many of our key AI assets, including Kunlun Chip and Apollo Go, reflect decade‑plus investment and are now core sources of differentiation whose performance and business value are increasingly recognized.ERNIE remains central to our AI strategy and full‑stack capabilities. We were among the earliest in China to invest in foundation models; while we iterated through mistakes, our commitment to ERNIE's competitiveness is unchanged. We will keep allocating the resources ERNIE needs, have optimized the org and added top AI talent, and are confident we can accelerate iteration and bring ERNIE back to the first tier.

Methodology will stay application‑driven. No single model can lead across all dimensions, so we focus ERNIE on capabilities that matter most to our own apps — AI Search, digital human, Miaoda, Pharmo, and general‑purpose agents like Doomate — with the apps themselves part of continuous improvement.In AI Search, as ERNIE advances in intent understanding and content quality assessment, improvements flow directly into search and feeds, yielding rapid impact and showing what still needs work. The related data then feeds model training, reinforcing ERNIE on these axes; this closed loop is a key path to converting tech progress into UX and real user and commercial value.

Q: What drives the sustained growth in AI cloud infra? What is the outlook for the next few quarters, and how do you view long‑term margin potential as the business scales?

A: AI cloud revenue grew 50% YoY in Q2, a solid pace above industry; over a longer window, cloud infra has posted record growth for multiple quarters and kept outgrowing peers. The highlight was GPU cloud +283% YoY, the fourth straight triple‑digit quarter and an acceleration vs. Q1's 184%.Looking ahead, several drivers stand out. First, China AI compute demand is very strong and will grow further as AI embeds deeper into real apps and workflows, especially as inference scales rapidly.

Second, the customer base is expanding fast with new adopters of various sizes, and existing core customers are lifting usage and spend. Third, use cases are broadening across Internet, gaming, embodied AI, AV, smartphones, and financial services.Fourth and most important, we have built a differentiated AI architecture spanning chips, cloud infra, models, and apps, with competitive products at each layer; our early bet at the app layer, including agents and products like Doomate, enhances our ability to capture diverse AI opportunities. Given demand trends, the pipeline, and these edges, we are confident AI cloud infra can stay strong in 2H, with room to accelerate.

On profitability, both profit and margin improved YoY in Q2. Three factors support further expansion: (1) GPU cloud is growing far faster than overall AI cloud infra, taking a larger mix with better margins than CPU cloud, and can improve further with scale, mix, utilization, and ops efficiency; (2) token revenue from external customers on Qianfan is rising very fast — still a small mix today, but as usage scales and unit inference cost keeps falling, MaaS can unlock larger profit contribution over time; (3) our AI stack and in‑house chips deliver end‑to‑end cost advantages. Net‑net, AI cloud infra has substantial long‑term margin upside.

Q: AI now makes up half of revenue and capex is ramping — how should we think about OPM trajectory? How do you balance ongoing AI investment with profitability?

A: AI‑related revenue remained about half of Baidu Core in Q2. Within that, AI cloud infra is growing revenue and profit rapidly with YoY margin gains; GPU cloud, with superior margins, continues to increase its revenue mix.Combined with strong market demand and cost advantages from in‑house chips and full‑stack AI, we see clear room for long‑term margin expansion, further supported by better utilization and operating leverage as scale builds.

AI apps also carry attractive long‑term profit potential — they are sticky and software‑heavy, and margins should improve with scale; as penetration rises and monetization matures, they should contribute more to overall profitability.We remain in an AI investment phase with firm commitment, but equally focus on spend where ROI is highest. Investments are driven by explicit customer and internal needs, so most spend is deployed quickly and begins contributing revenue earlier; we are also strengthening supply‑chain management to raise capital efficiency as we scale. Different AI investments monetize on different timelines, and some take longer to fully realize value; we will invest decisively where it most shapes long‑term competitiveness, while maintaining discipline on ROIC, operating efficiency, and cash flow. As AI scales and monetization matures, more of this spend should translate into sustainable profit growth.

Q: Progress and milestones for the Kunlun Chip spin‑off? What are its growth drivers, long‑term potential, and strategic role within BIDU's AI ecosystem?

A: The listing process is ongoing and we will update the market when we can. We are confident in Kunlun Chip's long‑term growth and commercial potential for several reasons.First, AI compute demand is rising for both training and inference, and as models advance and apps scale into real‑world scenarios — especially with agents expanding into more use cases — inference is accelerating, creating structural growth for AI chips.

Second, China's domestic market has significant upside while supply may remain tight for a period; customers increasingly need high‑performance, reliable, cost‑effective domestic AI chips. With decade‑plus investment, Kunlun Chip has strong performance, a tightly integrated soft‑hardware stack, compatibility with mainstream models/frameworks, and large‑scale deployment capabilities, with rising customer recognition and a favorable position to capture demand.Third, within BIDU's AI ecosystem, Kunlun Chip is a core part of the infra layer in a stack spanning chips, cloud, infra, models, and apps. Tight coordination across layers enables end‑to‑end optimization for better performance, reliability, and cost efficiency, supporting long‑term deployment of AI cloud infra and other AI businesses and strengthening full‑stack competitiveness. Looking ahead, Kunlun Chip will play a bigger role in BIDU's AI infra, addressing broader opportunities and more diverse demand.

Q: Timeline for dual primary listing in HK and potential Southbound inclusion? Strategic rationale and impact on investor base, liquidity, and valuation?

A: The BOD approved the conversion in Jul, application has been submitted and accepted by HKEX; next is the Aug 26 EGM to seek shareholder approvals on required items. Conversion is expected to take effect within the year, subject to HKEX approval and other applicable conditions.On Southbound inclusion, we are preparing actively post‑conversion and aim for early inclusion, subject to eligibility and exchange review processes and decisions.Strategically, dual primary aims to broaden the investor base, improve liquidity, and increase financing flexibility across HK and the U.S., while enabling more investors — especially in Asia — to understand and participate in BIDU as an AI‑first company. Longer term, if Southbound inclusion is achieved, Mainland investor participation should rise and the shareholder base should become more diversified over time.

Q: Progress on AI Search product, UX, and monetization? What drove ongoing pressure in online marketing in Q2, and how does 2H look?

A: Recent AI transition efforts focused on upgrading AI answer quality to improve UX. Accuracy and authority — BIDU's long‑standing strengths — have been reinforced; AI Search better understands user intent, with more reliable, better‑structured, and richer answers, while keeping hallucination low and improving content quality judgment, yielding more high‑quality and fewer low‑quality answers.Users are responding well, with satisfaction, intent to search, and retention improving steadily. We also further integrated AI Search with ERNIE Assistant to convert one‑off answers into coherent, interactive multi‑turn conversations.

We continue to strengthen multi‑step planning and complex task execution. ERNIE Assistant's task agent recently topped two influential benchmarks — a global list focused on real‑world task completion (transcript 'Pinch Bench V2') and a leading domestic agent evaluation (transcript 'ex‑cloud') — validating leadership in planning and execution.Competition remains intense as AI‑native product formats gain traction and user discovery/consumption habits evolve, intensifying the fight for time and attention. Meanwhile, we are deliberately holding back on monetization and ad load/pricing in AI Search during the transition, both of which have weighed on ads recently; these factors will likely persist, keeping ads under pressure in 2H. Monetization will follow product and UX — as model capability, UX, and task completion improve, more monetization hooks that fit naturally into AI experiences will emerge.

Q: Views on China's latest Robotaxi policies and the regulatory evolution? How do you balance domestic vs. overseas priorities and pace, and what progress has been made in overseas commercialization?

A: The global Robotaxi industry is moving quickly. The focus has shifted from safety and comfort to reliable, scalable ops and integration into broader transport systems; in step, major markets are iterating regulatory frameworks.In China, the first mandatory national standard for L3/L4 safety was released recently, to which Apollo Go contributed extensive tech and ops experience on the L4 requirements. Safety remains our top priority; we maintain a leading safety record globally and will hold a high bar on safety and operations. Clearer, more systematic regulation should raise industry standards, build public trust, and support orderly, long‑term growth.

Against this backdrop, we are positive on Apollo Go's global expansion and do not frame domestic vs. overseas as either‑or, but remain open and flexible by city. We assess each market's regulatory framework, mobility needs, pricing, road conditions, and commercial feasibility to set entry and scale‑up pace; with proven tech and ops, we can move fast where conditions allow, aiming to build deep presence in every city we enter regardless of borders.Specifically, Dubai has entered fully driverless commercial ops at growing scale and is the largest local provider; in London we are testing and co‑developing with Uber and Lyft; Hong Kong became the first right‑hand‑drive, left‑lane market with fully driverless testing; Shenzhen order volumes are rising fast and it is among our largest markets.

On unit economics, as fleets scale and the ops model matures, vehicle and ops costs should keep falling, with scale adding further efficiencies. We have achieved breakeven unit economics in markets with relatively low ride prices; in higher‑price overseas markets, low‑cost vehicles plus a mature ops model could drive even stronger unit economics. Intl markets outside the U.S. and China are larger than China's domestic market, offering substantial addressable opportunities, and we are confident in driving more cities to breakeven on the back of our tech, cost, and ops advantages.

Risk Disclosure & Statement:Dolphin Research Disclaimer and General Disclosure